Certifying Model Upgrades with Slice-Wise Non-Regression and Incumbent Fallback
An auditable protocol is established that separates candidate search from independent, paired evaluation and returns the exact incumbent when certification fails, and finite-sample guarantees for one frozen candidate, a finite candidate library, and a prespecified testing order are state.
Sheng-Wei Zhang, Tao Wu, Fei-Fei Qian
· 1 citation